Fig. 5.

Training dynamics for both splitting strategies. A key distinction between the training on the TCR and strict splits is that the validation performance steadily converges only in the TCR split case. In the strict split case, no significant improvement is achieved even after training for more than 140 epochs. This indicates that in the strict split scenario, training and validation data may be too distinct to enable a proper generalization. The AA CDR3 setting was used for this plot.